Vision-Based Estimation of MDS-UPDRS Gait Scores for Assessing Parkinson's Disease Motor Severity

Mandy Lu, Kathleen L. Poston, Adolf Pfefferbaum, Edith V. Sullivan, Li Fei-Fei, Kilian M. Pohl, Juan Carlos Niebles, Ehsan Adeli

TL;DR

This paper introduces a novel computer vision approach to automatically estimate Parkinson's Disease (PD) motor severity using gait analysis from video. By extracting 3D body skeletons and employing a hybrid ordinal-focal loss function, the method achieves high accuracy in classifying PD severity, offering a non-intrusive and objective biomarker for disease progression.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is a progressive neurological disorder that significantly impacts motor function. Accurately and quantitatively assessing the severity of these motor impairments is crucial for developing effective treatments and monitoring disease progression. Current methods for assessing PD severity often rely on subjective clinical ratings (like the MDS-UPDRS scale), expensive and intrusive wearable sensors, or neuroimaging techniques. The authors aim to bridge this gap by developing a non-intrusive, vision-based system that can objectively quantify PD motor severity, specifically focusing on gait impairments, using readily available video recordings.

Methods

The proposed framework, OF-DDNet, involves several key steps:

  • Participant Detection and Tracking: The system first identifies and tracks the PD patient within the video to isolate their movements from other individuals (e.g., clinicians). This is achieved using the SORT algorithm.
  • 3D Body Mesh and Pose Extraction: For each frame, the 3D pose (joint coordinates) of the participant is extracted. This is done using the SPIN (SMPL oPtimization IN the loop) method, which leverages a pre-trained SMPL body model and is fine-tuned on large public datasets. This step ensures that the pose estimation is robust and can be performed even with monocular video input.
  • Gait Score Estimation with OF-DDNet: The core of the method is a temporal convolutional neural network (TCNN) architecture, based on DD-Net, that takes the sequence of 3D poses as input. To address the challenges of imbalanced clinical datasets and the ordinal nature of the MDS-UPDRS scores, the authors introduce a hybrid ordinal-focal loss.
  • Focal Loss: This component is designed to combat class imbalance by down-weighting the contribution of easily classified examples, allowing the model to focus on harder-to-classify instances.
  • Ordinal Loss: This component leverages the inherent order of the MDS-UPDRS scores (0 being least severe, 4 being most severe) to penalize predictions that violate this order.

The combination of these losses aims to improve the model's performance on imbalanced datasets and its ability to accurately predict the ordered severity scores.

Datasets

  • Clinical Dataset: The primary dataset consists of video recordings from 30 research participants who met UK Brain Bank diagnostic criteria for PD. These videos were recorded during the "off-medication" state and were scored by a board-certified movement disorders neurologist using the MDS-UPDRS scale. The dataset has a significant class imbalance, with many more examples of lower severity scores (0 and 1) compared to higher severity scores (2 and 3).
  • CASIA Gait Database A: To augment the control cohort (score 0), samples from this public dataset of non-PD participants were included.

The authors introduce a new benchmark for classifying PD patients based on MDS-UPDRS gait severity using video.

Key Insights and Ideas

  • 3D Pose is Crucial: The study strongly suggests that 3D pose information provides significantly more valuable insights for PD severity prediction compared to 2D pose data, even for simpler baseline models.
  • Classification Outperforms Regression: For this task, classification-based methods demonstrated superior performance over regression-based approaches when using 3D joint data.
  • Hybrid Loss Effectiveness: The proposed hybrid ordinal-focal loss is key to handling the challenges of imbalanced clinical datasets and leveraging the ordered nature of the MDS-UPDRS scores, leading to improved accuracy.
  • Non-Intrusive Assessment: The work highlights the potential of computer vision to provide objective, non-intrusive, and scalable methods for monitoring motor impairments in PD patients, which can be valuable in remote or resource-limited settings.

Key Results

The OF-DDNet method achieved state-of-the-art performance, setting a new benchmark for this task. Key quantitative results include:

  • Macro-average F₁-score: 0.83
  • AUC: 0.90
  • Precision: 0.86
  • Balanced Accuracy (average recall): 81%

The confusion matrix indicates that the model performs best on well-represented classes (control and class 1) and struggles slightly with class 2, which is also noted as being difficult to distinguish in clinical practice. Comparisons with baseline and ablated methods showed that OF-DDNet consistently outperformed others, with statistical significance (p < 0.05).

Limitations and Future Work

The authors acknowledge several limitations:

  • Small Dataset Size: The relatively small clinical dataset carries a risk of overfitting and uncertainty in the results. They mitigated this through data augmentation and using simpler models.
  • Class Imbalance: While addressed by the hybrid loss, the inherent imbalance remains a challenge.
  • Subjectivity of MDS-UPDRS: The MDS-UPDRS scale, despite standardization efforts, still contains inherent subjectivity, leading to potential disagreements among physicians. This is particularly evident in distinguishing between scores 1 and 2.
  • Control Participant Shortage: Due to a lack of control participants in their clinical dataset, they relied on a public dataset, which, while normalized, might still introduce subtle differences.

Future work could involve expanding the clinical dataset, particularly by recruiting more participants from underrepresented classes, to further strengthen the results and improve generalization.

Summary updated

BibTeX
@inproceedings{lu2020vision,
  author = {Mandy Lu and Kathleen L. Poston and Adolf Pfefferbaum and Edith V. Sullivan and Li Fei-Fei and Kilian M. Pohl and Juan Carlos Niebles and Ehsan Adeli},
  title = {{Vision-Based Estimation of MDS-UPDRS Gait Scores for Assessing Parkinson's Disease Motor Severity}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  doi = {10.1007/978-3-030-59716-0\_61},
}

Loading related content...

Figure